---
id: "lib-2026-05-00-the-ai-infrastructure-triad-in-regional-governance"
type: "library"
title: "The AI infrastructure triad in regional governance"
lang: "EN"
org: "Harvard Kennedy School"
source_type: "Report"
date: "2026-05-01"
week: "2026-W18"
url: "https://nosov.ai/library/lib-2026-05-00-the-ai-infrastructure-triad-in-regional-governance/"
original_url: "https://www.hks.harvard.edu/sites/default/files/2026-05/The%20AI%20Infrastructure%20Triad%20in%20Regional%20Governance_%20How%20Regions%20Balance%20Progress%2C%20Sustainability%2C%20and%20Equity.pdf"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "A Harvard Kennedy School working paper accepted by the peer-reviewed journal AI & Society, built on a previously coded set of 10,068 policy comments, though its cases are illustrative rather than a causal test."
summary_basis: "full-text"
original_minutes: "45"
tags: ["ai infrastructure", "governance", "sustainability", "equity"]
---

# The AI infrastructure triad in regional governance

Tier A (Strong evidence): A Harvard Kennedy School working paper accepted by the peer-reviewed journal AI & Society, built on a previously coded set of 10,068 policy comments, though its cases are illustrative rather than a causal test.

Summary based on the full text.

## In short

A May 2026 working paper by Paulo Carvão (Senior Fellow, Mossavar-Rahmani Center for Business & Government, Harvard Kennedy School) and Tushar Kanade (MIT), accepted for publication in AI & Society. It argues that the "innovation versus regulation" framing misses the real choice regions face when hosting AI data centers: a trade-off between Progress (scale and speed), Sustainability (energy, water, land viability) and Equity (who gains, who pays, who decides). The authors say regions are unlikely to maximise all three at once. Where no deliberate choice is made, the balance defaults to whoever holds the most capital, which usually means Progress.

## Key numbers

- 10,068: public comments on the 2025 U.S. AI Action Plan used to map stakeholder worldviews
- 26%: share of Dominion Energy Virginia's electric load consumed by Northern Virginia data centers
- 40.6%: Big Tech submissions classed as Accelerationist
- PUE 1.25 / 50%: Singapore DC-CFA2 maximum power usage effectiveness and minimum green-energy share for new data centers

## What it reports

- Global data center electricity use was about 415 TWh in 2024 and is projected to more than double by 2030 (IEA); U.S. data centers use over 4% of national electricity.
- Northern Virginia facilities consume about 26% of Dominion Energy Virginia's load, with leasing vacancy below 1%; an estimated 20% of planned global projects face grid-related delays.
- Worldviews in the comment dataset: Big Tech is 40.6% Accelerationist, 37.5% National Security Hawk, 21.9% Responsible AI Advocate. Civil society is 43.7% Public Interest AI and 30.8% Responsible AI Advocate. Small tech splits 25.7% Accelerationist and 26.7% Responsible AI.
- Singapore froze new data centers from 2019 to 2022, then allocated about 80 MW through a pilot call. A second call (December 2025) allocates at least 200 MW with a 50% green-energy minimum and a 1.25 PUE cap.
- UAE and Saudi Arabia announced more than $100 billion in AI infrastructure ambitions, including the 1 GW Stargate UAE cluster and 500 MW of HUMAIN capacity.
- Texas SB 6 (June 2025) set cost-sharing and load-curtailment rules for customers above 75 MW.
- In January 2026 Microsoft pledged to pay full electricity costs and full property taxes. OpenAI's "Stargate Community" included a $1 billion equity investment in SB Energy for a 1.2 GW Texas campus.
- The paper ends with a five-question "Deliberate Triad Choice Framework": which dimension a decision advances, who decided, the explicit trade-offs, how costs and benefits are distributed, and what accountability exists.

## How they got it

The paper is conceptual. It combines earlier work on the economic, physical and moral limits of AI with a prior LLM-assisted coding of 10,068 Action Plan comments (GPT-4 Turbo and Gemini 1.5 Flash compared, with manual correction). It then reads five regional cases (Taiwan, Singapore, Gulf States, Northern Virginia, Texas) through the triad. The cases were chosen for variety, not representativeness.

## Limits

- The authors state that the evidence is illustrative and that the worldview-to-triad mapping is interpretive, not tested.
- The Microsoft and OpenAI commitments are self-reported pledges whose motives (branding, pre-empting regulation) the paper itself flags as open.

## Why it matters for companies

AI capacity is now constrained by power, land and local consent, and these costs reach electricity prices and permitting timelines for everyone in a region. Firms siting plants, offices or their own compute should expect data center competition for grid capacity. The five questions also work inside a company: any large AI investment can be tested for which objective it serves, who was consulted and who bears the cost.

Original: https://www.hks.harvard.edu/sites/default/files/2026-05/The%20AI%20Infrastructure%20Triad%20in%20Regional%20Governance_%20How%20Regions%20Balance%20Progress%2C%20Sustainability%2C%20and%20Equity.pdf